Live site  |  How it works  |  Architecture  |  Data  |  Run it locally

Corruption in Peru is already public. Vigía makes it readable. Every tender, contract and debarred company is open data. The bottleneck was never the data, it was the cost of reading it. Vigía funds that reading with small donations: a contract is read for S/ 3, and the result is public, always.

Investigating one suspicious state contract takes a journalist days: 80-page scanned PDFs spread over a dozen .gob.pe portals that do not talk to each other. Peru has published about 20 thousand contracts in 2026 alone. Oversight is a slow, heroic act that does not scale. Vigía puts AI agents to read the bidding file, cross it with official records and point at the contracts that deserve a second look, each signal linked to the law it cites and the document it comes from.

flowchart LR
    A["🏛️ OECE publishes<br/>a contract"] --> B["📥 Daily ingestion<br/>contract and documents"]
    B --> C[("🗄️ Postgres<br/>one source of truth")]
    D["🤝 Someone donates<br/>S/ 3 per contract"] --> E["🎲 Contracts assigned<br/>nobody chooses which"]
    C --> E
    E --> F["🤖 AI agent reads<br/>bases, minutes, contract"]
    F --> G{"⚖️ 4 judges<br/>check the result"}
    G -->|solid| H["✅ Published signal<br/>law and source attached"]
    G -->|doubtful| I["👀 Human review"]
    H --> J["🗺️ Map, rankings,<br/>donor receipt"]

Each reading ends with four self-evaluation judges (flag support, evidence citation, non-accusatory tone, pipeline completeness). A doubtful result goes to a person instead of being published, and a published result stays public even if it points at the donor who paid for it.

Every contract is linked to the party that governed the contracting entity in the month the contract was called, using the regional and municipal election results of 2014, 2018 and 2022 (JNE). A donor can fund the reading of the contracts of one party in one electoral term, and the contracts are picked at random among those entities.

flowchart LR
    J["📜 JNE results<br/>2014, 2018, 2022"] --> G["gobierno_periodo<br/>who governs each place,<br/>per term"]
    E["🏢 Entities<br/>by name: Regional Government,<br/>Provincial or District Municipality"] --> EG["entidad_gobierno"]
    G --> CP["contrato_partido"]
    EG --> CP
    C["📄 Contract<br/>call date"] --> CP
    CP --> R["📊 Ranking and funding<br/>per party and term"]

flowchart TB
    P(["👥 People"]) --> W
    subgraph CF["☁️ Cloudflare Workers"]
        W["vigia-web<br/>Next.js 14, OpenNext"]
        API["vigia-api<br/>Hono, TypeScript"]
        DSP["vigia-dispatcher"]
        MCP["vigia-mcp"]
        REL["vigia-relay<br/>Zurich"]
        W --> API
    end
    subgraph GCP["🟦 Google Cloud"]
        PGB["PgBouncer<br/>e2-micro VM"]
        SQL[("Cloud SQL<br/>Postgres 16 and PostGIS")]
        AG["🤖 Agents on Cloud Run<br/>one per contract type, Google ADK"]
        VX["Vertex AI<br/>Gemini 3.6 Flash"]
        DOC["Document AI"]
        RAG["RAG Engine<br/>4 corpora of law"]
        JOB["⏱️ Cloud Run Jobs<br/>and Cloud Scheduler"]
        PGB --> SQL
        AG --> VX
        AG --> DOC
        AG --> RAG
    end
    API -->|"Hyperdrive and tunnel"| PGB
    DSP --> AG
    AG --> PGB
    MCP --> PGB
    JOB --> SQL
    REL --> OECE
    JOB --> OECE
    subgraph SRC["🏛️ Public sources"]
        OECE["OECE, OCDS API"]
        PNDA["PNDA"]
        JNE["JNE"]
        MEF["MEF"]
        ONPE["ONPE"]
    end
    JOB --> PNDA
    JOB --> JNE
    JOB --> MEF
    AG -.->|traces| ARZ["📊 Arize Phoenix"]

The raw datasets are not in this repository: they are large and some contain personal data (the RNP registry has partners' names and IDs). We link the official open-data sources instead. Details in docs/DATA.md.

> show me the high-risk
  contracts in Áncash

each result carries:
  the law it cites
  the official document
  the state of the reading

A reading costs about US$ 0.17 after the Flex pricing and caching work, against a S/ 3 donation per contract.

OpenInference instruments the ADK runner and every Gemini call and exports to Arize Phoenix: a full span tree per contract with tokens, cost, latency, prompts and responses. If an agent drifts, it shows up in the trace.

vigia-peru/
├── frontend/                 Next.js 14: portada, /app (map, contracts, signals, entities, parties, funding), /admin
├── backend/
│   ├── agent/                ADK agents in Python, one per contract type, deterministic pipeline
│   ├── api/                  Read and write API (Hono) for Cloudflare Workers and Cloud Run
│   ├── dispatcher-worker/    Hands paid contracts to the agents
│   ├── mcp/ and mcp-worker/  Read-only MCP server
│   ├── relay-worker/         Zurich relay to the OECE API
│   ├── r2-sync-worker/       GCS to R2 file copy
│   ├── db/migrations/        Numbered SQL migrations (01 to 41) and apply_all.py
│   ├── scrapers/             Ingestion pipelines: PNDA, OCDS, MEF, ONPE, JNE
│   └── scripts/              Loaders and generators (who governs each place, README charts)
├── infrastructure/
│   ├── terraform/            The platform as code
│   └── deploy/               One script per component, and the project-move runbook
└── docs/                     DATA.md, DATASET_MAP.md, MARCO_NORMATIVO.md, images

Each component ships a .env.example.

# Database: Postgres 16 with PostGIS
docker compose -f infrastructure/docker-compose.yml up -d
python backend/db/apply_all.py              # PGHOST=127.0.0.1 PGPASSWORD=vigia by default

# API
cd backend/api && npm install && cp .env.example .env && npm run dev
npm test                                     # unit tests

# Web
cd frontend && npm install && cp .env.example .env.local && npm run dev

# Scrapers
cd backend && pip install -r scrapers/requirements.txt && python -m pytest scrapers/tests

# Regenerate the charts in this README from the live API
python backend/scripts/generar_graficos_readme.py

1. We never accuse anyone. A signal is a reason to review, never a charge.
2. We never publish citizens' personal data. Only serving officials and companies that contract with the State.
3. We do not replace the Comptroller, the prosecutors or journalism. We detect and prioritize.
4. Citizen reports: two independent reports within 30 days to be confirmed. No photo, not published.
5. Funding buys reading, not results: donors do not choose contracts, cannot edit findings, and are published even if flagged.
6. Every flag links to official evidence. No verifiable link, not published.
7. No ads, no data monetization. Open source and non-profit.

A working platform with national coverage of published contracts (about 20 thousand in the base), the three-reading map, funding by zone, entity and party, public donor profiles and an admin panel. Reading with AI is funded by donations, so the share of contracts read grows as people contribute.

Known gaps: ONPE contributions are loaded by hand, payments are validated manually (there is no gateway yet), and the 2026 election results will be added when the JNE proclaims them.

Spanish documents: ARQUITECTURA.md, SOLUCION.md, OBSERVABILITY.md.